Chinese robot manufacturers fund their operations with data sales to western competitors, heavily subsidized by local Chinese governments.Last year, China reportedly built and shipped over 10,000 humanoid robots.

The hardware is arguably ready, and the iteration speed is staggering. The major chinese OEMs release at least one new hardware model every twelve months, each one significantly better and cheaper than the previous generation.Even so, almost all of these robots are sold for research and development purposes, with very few having been put to actual productive work. The hardware is significantly more mature than the software.

To close this gap, analysts, researchers and investors currently seem to agree that there is a glaring bottleneck: training data.

While LLMs could be trained on the scraped corpus of the entire internet, there is vanishingly little training data for physical AI and robots. The last twelve months has seen an explosive increase in data collection efforts, and skyhigh valuations for AI labs with access to unique datasets.

This weekend, I was invited to visit a new kind of robotics factory - one that didn't produce robots, or even hardware components, but training data.Occupying two whole floors of a newly built high-rise in a massive new industrial park in Changzhou, entirely sponsored by the local government, 150 semi-humanoid robots had been set up to record up to a thousand hours of training data per day.

The data is resold to frontier AI labs all over the world for an average of 90 USD per hour, where the data is used to train better models. According to the data factory, it takes their customers as little as 200 hours of training data to train for a simple scenario. On the day we visited they were collecting data for over 250 distinct tasks.

Realman is not the only robotics OEM to run a data factory. AGIBot, who are neck and neck with Unitree in the domestic robotics race, tell us they made more money in the West last year selling data than selling robots - even though they're the secret white label manufacturer behind at least three well-known western robotics brands valued cumulatively at over a billion dollars.When we visited their data factory in Shanghai in the summer of last year they had several dozen robots teleoperated in a wide variety of scenarios, with customers ranging in market cap from trillions to millions.

The unit economics are tantalizing. The typical data collection robot retails for about 50,000 USD, but the OEMs themselves have a significantly lower manufacturing cost. While no one has been willing to go on the record to confirm their bottom line, it seems likely that several of them are already hovering around 10,000 USD per robot manufactured - with some rumored to already have broken into the four digits.

The data collection labor in places like Changzhou is roughly 3000-4000 RMB per month (around 500 USD). An hour of data costs them on the order of 5-10 dollars to produce. Assuming no overtime, each robot can produce an alleged baseline of 80 hours of data per month - but realistically, with 996 hustle they're producing over 100 hours of data per robot per month.It's seems likely that with current demand for data, these robots hit positive ROI within a quarter of a year for the OEM themselves - and well under a year for data factories that pay the full 50,000 USD per robot.

While this market opportunity may not exist forever, several OEMs are rapidly expanding their data collection efforts, and almost always with very attractive government subsidies. Realman is planning to open five additional factories across China this year, with support from local governments, each one deploying an additional 100 robots or more.

While the claim is that they can produce 30 minutes of data per hour of labor, their hallway dashboard indicates that while they're currently producing several hundred hours per day, they're not yet running at their projected peak capacity.

Threats to the business model include an emerging preference for egocentric data over teleoperation data, but the data factories have been quick to respond. Mass produced egocentric headbands and glasses, as well as UMI grippers, have already hit the domestic Chinese market - and prices are collapsing in real-time. Earlier this year it cost 15,000 USD to buy a premium UMI data collection setup, but one OEM revealed they have a new product line launching in September-October that will sell for as little as 2,000 USD.

At least one western company has pre-ordered over 1000 pairs to accelerate their internal data capture efforts. Other frontier labs are starting to bet heavily on synthetic data from simulations, which they argue is the only realistic way to close the 100,000 year gap.

There's also growing disagreement (and disillusionment) about the concept of data flywheels. @KyleVedder of Physical Intelligence recently argued that the data collection challenge is misunderstood by many, and that much of the data being produced (and about to be produced) will actually have very limited value.In summary, much has been said already about the fact that most robotics research is done on Chinese hardware today, and there's fiery disagreement about if the proposed Guard Act will help or hinder western robotics acceleration. Less discussed is that the Chinese OEMs are funding their development by selling data to the western labs.If data truly is the moat we must ask ourselves: why are they selling?

‍

Authored by Nils Pihl, CEO @ Auki Labs

‍

アウキ・ラボについて

Aukiはポーズメッシュという地球上、そしてその先の1000億の人々、デバイス、AIのための分散型機械認識ネットワークを構築しています。ポーズメッシュは、機械やAIが物理的世界を理解するために使用可能な、外部的かつ協調的な空間感覚です。

私たちの使命は、人々の相互認知能力、つまり私たちが互いに、そしてAIとともに考え、経験し、問題を解決する能力を向上させることです。人間の能力を拡大させる最も良い方法は、他者と協力することです。私たちは、意識を拡張するテクノロジーを構築し、コミュニケーションの摩擦を減らし、心の橋渡しをします。

ポーズメッシュについて

ポーズメッシュは、分散型で、ブロックチェーンベースの空間コンピューティングネットワークを動かすオープンソースのプロトコルです。

ポーズメッシュは、空間コンピューティングが協調的でプライバシーを保護する未来をもたらすよう設計されています。いかなる組織の監視能力も制限し、空間のプライベートな地図の自己所有権を奨励します。

分散化はまた、特に低レイテンシが重要な共同ARセッションにおいて、競争優位性を有します。ポスメッシュは分散化運動の次のステップであり、成長するテック大手のパワーに対抗するものです。

アウキ・ラボはポスメッシュにより、ポーズメッシュのソフトウェア・インフラの開発を託されました。